Improper Neutralization of Input Used for LLM Prompting

Opis

The product uses externally-provided data to build prompts provided to large language models (LLMs), but the way these prompts are constructed causes the LLM to fail to distinguish between user-supplied inputs and developer provided system directives.

When prompts are constructed using externally controllable data, it is often possible to cause an LLM to ignore the original guidance provided by its creators (known as the "system prompt") by inserting malicious instructions in plain human language or using bypasses such as special characters or tags. Because LLMs are designed to treat all instructions as legitimate, there is often no way for the model to differentiate between what prompt language is malicious when it performs inference and returns data. Many LLM systems incorporate data from other adjacent products or external data sources like Wikipedia using API calls and retrieval augmented generation (RAG). Any external sources in use that may contain untrusted data should also be considered potentially malicious.

Inne nazwy

prompt injection

attack-oriented term for modifying prompts, whether due to this weakness or other weaknesses

Skutki

• Confidentiality, Integrity, Availability

  Execute Unauthorized Code or Commands; Varies by Context

  The consequences are entirely contextual, depending on the system that the model is integrated into. For example, the consequence could include output that would not have been desired by the model designer, such as using racial slurs. On the other hand, if the output is attached to a code interpreter, remote code execution (RCE) could result.

• Confidentiality

  Read Application Data

  An attacker might be able to extract sensitive information from the model.

• Integrity

  Modify Application Data; Execute Unauthorized Code or Commands

  The extent to which integrity can be impacted is dependent on the LLM application use case.

• Access Control

  Read Application Data; Modify Application Data; Gain Privileges or Assume Identity

  The extent to which access control can be impacted is dependent on the LLM application use case.

Zapobieganie

• Architecture and Design

  LLM-enabled applications should be designed to ensure proper sanitization of user-controllable input, ensuring that no intentionally misleading or dangerous characters can be included. Additionally, they should be designed in a way that ensures that user-controllable input is identified as untrusted and potentially dangerous.

  Skuteczność: High

• Implementation

  LLM prompts should be constructed in a way that effectively differentiates between user-supplied input and developer-constructed system prompting to reduce the chance of model confusion at inference-time.

  Skuteczność: Moderate

• Architecture and Design

  LLM-enabled applications should be designed to ensure proper sanitization of user-controllable input, ensuring that no intentionally misleading or dangerous characters can be included. Additionally, they should be designed in a way that ensures that user-controllable input is identified as untrusted and potentially dangerous.

  Skuteczność: High

• Implementation

  Ensure that model training includes training examples that avoid leaking secrets and disregard malicious inputs. Train the model to recognize secrets, and label training data appropriately. Note that due to the non-deterministic nature of prompting LLMs, it is necessary to perform testing of the same test case several times in order to ensure that troublesome behavior is not possible. Additionally, testing should be performed each time a new model is used or a model's weights are updated.

• Installation, Operation

  During deployment/operation, use components that operate externally to the system to monitor the output and act as a moderator. These components are called different terms, such as supervisors or guardrails.

• System Configuration

  During system configuration, the model could be fine-tuned to better control and neutralize potentially dangerous inputs.

Przykłady podatności

• CVE-2023-32786 Chain: LLM integration framework has prompt injection (CWE-1427) that allows an attacker to force the service to retrieve data from an arbitrary URL, essentially providing SSRF (CWE-918) and potentially injecting content into downstream tasks.
• CVE-2024-5184 ML-based email analysis product uses an API service that allows a malicious user to inject a direct prompt and take over the service logic, forcing it to leak the standard hard-coded system prompts and/or execute unwanted prompts to leak sensitive data.
• CVE-2024-5565 Chain: library for generating SQL via LLMs using RAG uses a prompt function to present the user with visualized results, allowing altering of the prompt using prompt injection (CWE-1427) to run arbitrary Python code (CWE-94) instead of the intended visualization code.
• CVE-2024-48746 AI-based integration with business intel dashboard allows prompt injection through its natural language component, allowing execution of arbitrary code

Słabość nadrzędna

• CWE-77 Improper Neutralization of Special Elements used in a Command ('Command Injection')

Kiedy powstaje

• Architecture and Design — LLM-connected applications that do not distinguish between trusted and untrusted input may introduce this weakness. If such systems are designed in a way where trusted and untrusted instructions are provided to the model for inference without differentiation, they may be susceptible to prompt injection and similar attacks.
• Implementation — When designing the application, input validation should be applied to user input used to construct LLM system prompts. Input validation should focus on mitigating well-known software security risks (in the event the LLM is given agency to use tools or perform API calls) as well as preventing LLM-specific syntax from being included (such as markup tags or similar).
• Implementation — This weakness could be introduced if training does not account for potentially malicious inputs.
• System Configuration — Configuration could enable model parameters to be manipulated when this was not intended.
• Integration — This weakness can occur when integrating the model into the software.
• Bundling — This weakness can occur when bundling the model with the software.

Dotyczy

Not Language-Specific, Not OS-Specific, Not Architecture-Specific, AI/ML

Źródła

1. OWASP Top 10 for Large Language Model Applications - LLM01 — OWASP, 2023
2. IBM - What is a prompt injection attack? — Matthew Kosinski, Amber Forrest, 2024
3. Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection — Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, Mario Fritz, 2023